Not to dismiss this, but when I was working on a problem on Windows recently my LLM (6.1 Sol / oh-my-pi) ended up doing the same thing, although it used Qemu and snapshots to accomplish that.
Eventually it got to a point it could iterate very quickly and was really only constrained by total system RAM which meant a dozen at most copies could be running simultaneously.
Eventually it got to a point it could iterate very quickly and was really only constrained by total system RAM which meant a dozen at most copies could be running simultaneously.